Two Separate Continually Online Trained
Neurocontrollers for a Unified Power Flow Controller
Radha P. Kalyani, Student Member IEEE, Ganesh K. Venayagamoorthy, Senior Member IEEE
Department of Electrical and Computer Engineering
University of Missouri- Rolla
Rolla, MO 65409, USA
rpk5f9@umr.edu and gkumar@ieee.org
electric energy source. Practically, these two devices are two
Voltage Source Inverters (VSIs), one connected in shunt with
the transmission line through a coupling transformer and the
other is inserted in series with the transmission line through an
interface transformer. The UPFC, by means of angularly
unconstrained series voltage injection, is able to control,
concurrently or selectively, the transmission line voltage,
impedance, and angle or, alternatively, the real and reactive
power flow in the line. The UPFC may also provide
independently controllable shunt-reactive compensation.
Neural networks are suitable for multi variable applications
as they can easily identify the interactions between the
system’s inputs and outputs. Their ability to learn and store
information about system nonlinearities allows neural
networks to be used for modeling and designing intelligent
controllers for power systems [4, 5]. Thus, offering
alternatives for conventional linear and nonlinear control. A
radial basis function neural network controller for a UPFC
based on the direct adaptive control scheme has been reported
to improve the transient stability performance of a power
system [6]. It is known that indirect adaptive control is able to
control a nonlinear system with fast changing dynamics, like
the power system better, since the dynamics are continually
identified by a model. Advantages of the neurocontrollers over
the conventional controllers are that they are able to adapt to
changes in the system operating conditions automatically
unlike the conventional controllers whose performances
degrade for such changes and are required to be re-tuned to
give the desired performance.
This paper presents the design of neurocontrollers (NCs) to
control the UPFC and the power system in a single machine
infinite bus power system setup. The design of shunt and
series continually online trained (COT) NCs are based on the
indirect adaptive control scheme to replace the existing PI
controllers in the shunt and the series branches of a UPFC. In
addition, two other neural networks called neuroidentifiers
(NIs) are designed to identify the hybrid complex nonlinear
dynamics of the UPFC and the power system. The
neurocontrollers are designed based on the dynamics modeled
by the neuroidentifiers. In all, four COT neural networks are
used for the complete UPFC control. Comparison of the
neurocontrollers and PI controllers’ performances for damping
system oscillation and voltage regulation are presented for
different operating points.
Abstract— The crucial factor affecting the modern power systems
today is load flow control. The Unified Power Flow Controller
(UPFC) provides an effective means for controlling the power
flow and improving the transient stability in a power network.
The UPFC has fast complex dynamics and its conventional
control is based on a linearized model of the power system. This
paper presents the design of neurocontrollers that control the
power flow and regulates voltage along a transmission line. Two
separate neurocontrollers are used for controlling the UPFC, one
neurocontroller for the shunt inverter and the other for the series
inverter. Simulation studies carried out in the PSCAD/EMTDC
environment is described and results show the successful control
of the UPFC and the power system with the two neurocontrollers.
Performances of the neurocontrollers are compared with the
conventional PI controllers for system oscillation damping.
Adaptive
control;
Neuroidentifier;
KeywordsNeurocontroller; Power system; Unified Power Flow
Controller (UPFC)
I.
INTRODUCTION
With the ever increasing complexities in power systems
across the globe and the growing need to provide stable,
secure, controlled, economic, and high-quality electric power
–especially in today’s deregulated environment – it is
envisaged that Flexible AC Transmission System (FACTS)
controllers are going to play a critical role in power
transmission systems [1]. Transmission congestion results
when there is insufficient capacity to transmit power over
existing lines and maintain the required safety margins for
reliability. FACTS enhance the stability of the power system
both with its fast control characteristics and continuous
compensating capability. The two main objectives of FACTS
technology are to control power flow and increase the
transmission capacity over an existing transmission corridor
[2].
Gyugyi proposed the Unified Power Flow Controller
(UPFC) a new generation of FACTS devices in 1991 [3]. The
UPFC is a combination of a static synchronous compensator
(STATCOM) and a static synchronous series compensator
(SSSC) which are coupled via a common dc link, to allow
bidirectional flow of real power between series output
terminals of the SSSC and the shunt terminals of the
STATCOM, and are controlled to provide concurrent real and
reactive series line compensation without an external energy
0-7803-7883-0/03/$17.00 © 2003 IEEE
308
Governor
Pref
Σ
1
∆ω
Turbine Synchronous
Generator
Simulator
2
UPFC
V1
Z1
V2
Z1
Rs1, Ls1
Exciter
AVR
Σ
Pout,
Qout
Shunt
Converter
Vdc
Series
Converter
Infinite
Bus
V1ref
+
V1ref
V1
Σ
Shunt
Verr Converter
Control
Series
Qerr
Converter
Control
Vdcerr
Perr
Q ref
Qinj
Σ
Σ V
dc
Σ
Pinj
Vdcref
Pref
Fig. 1 Single machine infinite bus system with the UPFC (the ‘plant’).
II.
UPFC in the dq reference frame are given in [7]. The
conventional shunt and series branch control of the UPFC is
briefly described below.
POWER SYSTEM WITH UPFC
For identifying and controlling the dynamics of a UPFC and
the power system, a single machine infinite-bus (SMIB) power
system is setup (Fig. 1) in PSCAD/EMTDC environment. This
power system comprises of a synchronous generator with
exciter-automatic voltage regulator (AVR) and turbinegovernor combinations connected to an infinite bus through
two sections of transmission lines. The UPFC is placed
between the two sections of the transmission lines between
bus1 and 2 as shown in Fig. 1. The SMIB with the UPFC is
called the plant in the rest of this paper.
The series inverter provides the main function of the UPFC
by injecting a voltage with magnitude V2, which is controllable
and a phase angle δ in series with the line via an insertion
transformer. This injected voltage acts essentially as a
synchronous ac voltage source. The transmission line current
flows through this voltage source resulting in a reactive and
active power exchange between itself and the ac system. The
inverter generates the reactive power exchanged at the ac
terminal internally. The active power exchanged at the ac
terminal is converted into dc power, which appears at the dc
link as a positive or negative real power.
The basic function of shunt inverter is to generate or absorb
the real power demanded by series inverter at the common dc
link. The power demand by the series inverter at the dc link is
converted back to ac by the shunt inverter and fed to the
transmission line bus via a shunt-connected transformer. In
addition to this, the shunt inverter can also generate or absorb
controllable reactive power if desired and thereby provides
independent shunt reactive compensation for the line [3].
The three main control parameters of UPFC are voltage
magnitude, voltage angle and shunt reactive current. Control
of real and reactive power can be achieved by injecting series
voltage with an appropriate magnitude and angle. The
transient stability model for the shunt and series branch of a
A. Series Branch Control
The block diagram of the conventional PI controllers for
series branch of the UPFC is shown in Fig. 2. The control of
series inverter can be achieved using PQ-decoupled control.
The outputs of the control system are the modulation index k2
and phase shift α2. Neglecting the inverter losses, the injected
active power Pinj, reactive power Qinj, output active power Pout,
and reactive power Qout are given by the equations below.
Pinj =
V2 ( Eq − Eq cosδ + Ed sin δ )
X
V Ed cosδ + V2 Eq sin δ − V2 Ed + Ed 2 + Eq 2
Qinj = 2
X
Pout =
V2 2 sin δ + V2 E q
X
(1)
(2)
(3)
2V 2 E d cos δ + 2V 2 E q sin δ + E d 2 + E q 2
2X
(4)
V2 = Ed 2 + Eq 2 , Eq = Vinj sin(θinj ) , Ed = Vinj cos (θinj )
(5)
Q out =
where
It can be seen from (3) that Pout is mainly affected by
E q whereas (4) shows that Qout is affected by both E q and E d .
In incremental form, the line active and reactive power can be
expressed in terms of ∆E q and ∆E d as given by (6a) and (6b).
∆Pout =
309
V
∆E q
X
(6a)
shunt branch of UPFC (with the switches S1 and S2 at
position 1) is as shown in Fig. 3. The PI controllers are
replaced by the NC with switches S1 and S2 at position 3. The
design procedure of the NC is explained in section IV.
1
(∆EdV cos δ + ∆EqV sin δ + ∆Ed Ed 0 + ∆Eq Eq0 ) (6b)
X
But it can be assumed in practice that cos δ is close to unity
and sin δ is close to zero since the phase angle between the
two buses (receiving and sending ends) on a transmission line
is less than 30°, which leads to (7).
1
∆Qout = (V ∆ Ed + Edo ∆ Ed + Eqo ∆Eq )
(7)
X
The conventional PI control for the series branch of the UPFC
(with the switches S1 and S2 at position 1) is shown in Fig. 2.
The NI is trained with switches S1 and S2 at position 2 and the
NC controls the plant with switches S1 and S2 at position 3.
The design of NI and NC are explained in sections III and IV.
∆Qout =
III.
Two neuroidentifiers, one for the shunt inverter and the other
for the series inverter are used to identify the hybrid dynamics
of the UPFC and the power system. These networks
dynamically identify the controlling parameters of UPFC
∆ Epd, ∆ Epq, and ∆Ed, ∆Eq which are the outputs of the PI
controllers (Figs. 2 and 3). The NIs are developed using the
series-parallel Nonlinear Auto Regressive Moving Average
(NARMA) model [4]. The two neuroidentifiers are continually
online trained simultaneously to provide a dynamic model at all
times. This training takes place in two phases, namely a precontrol phase and a post-control phase [5].
2
Kp3 +
Qinj
Qref
∑
K I 3 ∆Eq
Qerr
Pref
∑
Ed 0
∆E q
3
TDL
TDL
∆ E d _ prbs
1
T3
∑
S1
K p4 +
∑
S2
3
Pinj
Vd c
Eq 0
∆E d
K I 4 ∆Ed
1
T4
A. Precontrol Phase
During this phase, the switches S1 and S2 in Figs. 2 and 3
are at position 2. The inputs to the NIs in this phase are the
outputs from plant and the pseudorandom random binary
signals (PRBS).
1) Series Neuroidentifier: The series UPFC branch
neuoidentifier (SENI) in Fig. 4 is a three layer feedforward
neural network with thirteen inputs, a single hidden layer with
fifteen sigmoidal neurons and two outputs. There are four
different types of inputs, the first two types are the differences
between the following signals: the measured real power and its
reference value - Perr, and, the measured reactive power and its
reference value Qerr. The other two types are the training
signals - ∆Ed_prbs and ∆Eq_prbs (Fig. 5). In the pre-training phase,
PRBSs are applied to excite all possible dynamics of the plant
[4, 5]. These PRBS are fed to the series inverter at B and SENI
at C with the switches S1 and S2 at position 2 (Fig. 2).
Ed 2 + Eq 2
k2 =
Neuro
Controller
Perr
α 2 = c o s−1
Ed
Ed 2 + Eq 2
∆ Eq _ prbs
2
Series Inverter
PWM
Fig. 2 Series inverter control showing conventional PI controllers, training
signals for the neuroidentifier (SENI) and neurocontroller (SENC).
B. Shunt Branch Control
Control of the shunt active and reactive current is achieved
by varying the shunt inverter voltage active Epd and reactive
components Epq respectively. The reactive power flow and
shunt input voltage can be regulated by active voltage
component Epd and the dc-link capacitor voltage Vdc support
can be achieved by regulating Epq. Fig. 3 shows a typical block
diagram of the conventional PI controllers for the UPFC’s
shunt branch control.
∆ E d _ prbs , ∆ E q _ prbs
∑
K I 1 ∆E pd
1
T
F
∆ E pd _ prbs
3
TDL
Vdcref
S1
∑
k 1=
E pq 0
∆E pq
S2
3
Vdc
K p2 +
KI2
T2
∆E pq
∑
2
Pˆ err (t ),
Qˆ err (t )
E p d2 + E p q2
Vd c
α 1= co s−1
Fig. 4 Pre-training of SENI to update the weights of the NI by
backpropagating the error signals at F.
Epd
E p d2 + E p q2
x 10-1
∆ E pq _ prbs
1
E
Series
Series
Neuroidentifier
Neuroidentifier
C
∑
Neuro
Controller
Vdcerr
Error
SENI
E pd 0
∆E pd
TDL
D
TDL
1
Verr
P err (t ), Q err (t )
TDL
TDL
PRBS Signals (p.u)
Vref
Kp1+
PLANT
B
2
V1
DESIGN OF NEUROIDENTIFIERS
Shunt Inverter
PWM
4
2
0
-3
-5
Fig. 3 Shunt inverter control showing conventional PI controllers, training
signals for the neuroidentifier (SHNI) and neurocontroller (SHNC).
5
The outputs of this control system are the modulation index
k1 and phase shift α1. The conventional PI control for the
10
15
Time (sec)
20
25
Fig. 5 Ep_prbs (one of the PRBS signals) applied to the series branch of UPFC
as shown in Fig. 2.
310
inverter at C and plant at B at during the pre-training phase
with the aid of switches S1 and S2 (Fig. 3). The outputs of
SHNI at E are compared to outputs of the plant at D and the
error signals at F are used to update the weights of the SHNI
using the backpropagation algorithm. This process is repeated
until a satisfactory error goal is obtained with the SENI
learning over a number of different possible operating points
of the plant.
Typical PRBS signal applied is shown in Fig. 5. The
frequency content of this signal is 1Hz, 3Hz and 5Hz. The
high magnitude of perturbation is required to have better
identification of the system dynamics. All the four different
types of inputs are time delayed (TDL) by one sample period
and together with their eight previously delayed values form
the twelve inputs to the SENI. The outputs of the SENI at E
are the estimated difference in the real power - P̂ err and in the
reactive power - Q̂ err at the next time step.
The outputs of SENI at E are compared to outputs of the
plant at D and the error signals at F are used to update the
weights of the SENI using the backpropagation algorithm. This
process is repeated until a satisfactory error goal is obtained
with the SENI learning over a number of different possible
operating points of the plant.
2) Shunt Neuroidentifier: The shunt UPFC branch
neuroidentifier (SHNI) in Fig. 6 is a three layer feedforward
neural network with thirteen inputs, a single hidden layer with
eighteen sigmoidal neurons and two outputs. There are four
different types of inputs, the first two inputs to the NI are
namely, the deviation signals between the measured shunt
voltage and its reference value Verr, the measured dc link
voltage and its reference Vdcerr, and the other two types are the
PRBS training signals ∆Epd_prbs and ∆Epq_prbs (Fig. 7) with
magnitudes in proportion to the real and reactive components
of shunt inverter voltage Epd and Epq respectively.
∆ E pd _ prbs , ∆ E pq _ prbs
PLANT
B
B. Postcontrol Phase
During this phase, online training of the SENI and SHNI
continue while the neurocontrollers are controlling their
respective branches of the UPFC, with switches S1 and S2 at
positions 3 (Figs. 2 and 3). The design of the neurocontrollers
is described in section IV. The PRBS signals used in the precontrol phase are now set to zero and outputs from the NCs are
applied to the plant. The post-training steps for NIs (Figs. 8
and 9) are described below.
A
PLANT
PLANT
B
TDL
Series
Series
Neurocontroller
Neurocontroller
SENC
D
Vˆerr (t ),
(t )
Vˆ
E
Series
Neuroidentifier
SENI
J
Fig. 8 Post-training of SENI to update the weights of the NI by
backpropagating the error signals at F.
Error
SHNI
Series
C
Shunt
Neuroidentifier
Neuroidentifier
dcerr
PLANT
PLANT
B
TDL
A
PRBS signals (V)
0.05
0.03
Shunt
Series
Neurocontroller
Neurocontroller
SHNC
Error
∆ E pq
F
TDL
20
25
30
K
∆ u (t ) =
∆ E pd ,
C
15
Time
Vdcerr (t )
TDL
TDL
0.01
0
-0.01
-0.03
D
Verr (t ),
Fig. 6 Pre-control training of SHNI for plant identification.
10
E
C
F
Pˆerr(t),
ˆ (t)
Q
err
F
Verr (t ), Vdcerr (t )
TDL
Qerr(t)
K
Error
∆E q
TDL
5
Perr(t),
TDL
TDL
∆ u (t ) =
∆ E d,
TDL
-0.05
D
E
Vˆerr (t ),
Vˆ
(t )
dcerr
Shunt
Neuroidentifier
SHNI
J
Fig.9 Post-training of SHNI to update the weights of the NI by
backpropagating the error signals at F.
Fig. 7 Epd_prbs (one of the PRBS signal) applied to the shunt branch of UPFC
as shown in Fig. 3.
i.
The plant output signals at D are sampled and time
delayed by one, two and three sample periods.
ii. The sampled signals from step i above are fed at A to the
NCs which then calculates the control signals ∆ Epd ,
∆ Epq, (SHNC) and ∆Ed, ∆Eq(SENC) which are applied to
control the plant.
All the four types of inputs are time delayed by one sample
period and together with their eight previously delayed values
form the twelve inputs to the SHNI at C (Fig. 6). The outputs
of the SHNI at E are the shunt voltage deviation Vˆ
and dc
err
link voltage deviation Vˆdcerr which are estimated one time
step ahead. These PRBS signals are only fed to the shunt
311
from the output of the desired response predictor at K to
produce the error signal at H which is backpropagated through
the SHNI to obtain desired control signal ∆ uˆ (t ) . The
difference between ∆ uˆ (t ) and the outputs of SHNC generates
the error signal at L which is used to update the weights of the
SHNC using backpropagation. Pre-training is terminated when
the weights of the SHNI and SHNC have converged over a
period of time at different operating points. The next phase of
the training (post-control) for the SHNC is carried out while the
SHNC is allowed to control the plant.
iii. These control signals are time delayed by one, two and
three sample periods, and, together with the signals from
step i are fed to the NIs at C.
iv. The outputs at D (Perr (t), Qerr (t) of series branch and
Verr(t), Vdcerr(t) of shunt branch) and the outputs of NIs at
E ( Pˆ (t ) , Qˆ (t ) of SENI and Vˆ (t ) , Vˆ
(t) of
err
err
err
dcerr
SHNI) are subtracted respectively to produce error signals
at F which are backpropagated to update weights of the
respective NIs.
IV. DESIGN OF NEUROCONTROLLERS
In the series and shunt UPFC branch neurocontroller design,
each consists of two separate neural networks, one for the
identifier/model (described in section III) and other for the
controller. The neurocontroller is used to replace the
conventional PI controllers in each branch (Fig. 3 and 4). The
training of neurocontrollers like the neuroidentifiers also takes
place in two phases, namely a pre-control phase and a postcontrol phase [5]. Both neurocontrollers are trained
simultaneously.
A. Precontrol Phase
During the precontrol phase the inputs to the NCs are the
perturbated outputs from the plant as shown in Figs. 10 and
11. The PRBS signals which are inputs to the NIs are also fed
to the plant to cause the necessary perturbations.
1)
Series Neurocontroller: The series UPFC branch
neurocontroller (SENC) in Fig. 10 is a three layer feedforward
neural network with six inputs, a single hidden layer with
eighteen sigmoidal neurons and two outputs. There are two
types of inputs to the SENC namely the Perr and the Qerr.
These signals at time t-1, t-2 and t-3 form the six inputs. The
two outputs of SENC (∆Ed and ∆Eq) are the control signals
∆u(t). The PRBS signals are applied to the plant and the SENI
to carry out the pre-training. The outputs of the plant are fed
into the desired response predictor [4], which predicts
Perr(t+1) and Qerr(t+1) at K. The output of SENI at E is
subtracted from the output of the desired response predictor at
K to produce the error signal at H which is backpropagated
through the SENI to obtain desired control signal ∆ uˆ (t ) . The
difference between ∆ uˆ (t ) and the outputs of SENC ∆u (t )
generates the error signal at L which is used to update the
weights of the SENC using backpropagation. Pretraining is
terminated when the weights of the SENC have converged
over a period of time for the PRBS signal applied over a
number of operating points of the plant.
2) Shunt Neurocontroller: The series UPFC branch
neurocontroller (SHNC) in Fig. 11, is a three layer
feedforward neural network with six inputs, a single hidden
layer with eighteen sigmoidal neurons and two outputs. Fig.
11 shows the SHNC development architecture and, the
respective inputs and outputs for the pretraining phase.
The PRBS signals are applied to the input of the shunt
UPFC branch and the SHNI. The outputs of the plant are fed
into the desired response predictor, which predicts Verr(t+1)
and Vdcerr(t+1) at K. The output of SHNI at E is subtracted
E d _ prbs , E q _ prbs
Series
Series
Neurocontroller
SENC
A
PLANT
B
TDL
Desired
Response
Predictor
P err (t + 1),
Q err (t + 1)
K
D
TDL
∆ u (t ) =
∆ Ed ,
Error
∆Eq
TDL
TDL
Error
∆ uˆ (t )
err
E
C
L
Pˆ err (t + 1),
Qˆ
(t + 1)
H
Series
Series
Neuroidentifier
Neuroidentifier
SENI
J
Fig. 10 Pre-training of SENC to update the weights of the NC by
backpropagating the error signals at L.
E pd _ prbs , E pq _ prbs
TDL
TDL
A
PLANT
PLANT
B
TDL
TDL
Shunt
Series
Neurocontroller
SHNC
D
∆ u (t ) =
∆ E pd ,
Desired
Desired
Response
Predictor
Predictor
Verr (t + 1),
V
(t + 1)
K dcerr
Error
∆ E pq
C
L
TDL
TDL
Error
Error
∆ uˆ (t )
E
Vˆerr (t + 1),
Vˆ
(t + 1)
dcerr
H
Shunt
Series
Neuroidentifier
Neuroidentifier
SHNI
J
Fig. 11 Pre-training of SHNC to update the weights of the NC by
backpropagating the error signals at L.
B. Postcontrol Phase
During this phase, online training of the NCs continue
while NCs are controlling their respective branches of the
UPFC. The PRBS signals used in the pre-control phase are now
set to zero and outputs from the NCs are applied to the plant
(with switches S1 and S2 in Figs. 2 and 3 at position 3). The
following steps are carried out during the post-control phase
(Figs. 12 and 13):
i. In the post-training of NCs, the output of the NIs at E
( Pˆ (t + 1) , Qˆ (t + 1)
of SENI and Vˆ (t + 1) ,
err
err
err
Vˆdcerr(t +1) of SHNI), and the desired response at K (Perr
(t+1), Qerr (t+1) of series branch and Verr(t+1), Vdcerr(t+1) of
312
tuned for this operating point using time response analysis [10].
A sampling frequency of 10 kHz is used to sample the outputs
of the plant.
shunt branch) are subtracted respectively to produce second
error signals at H. The error signals at H are backpropagated
through the NIs and their derivatives are obtained at J (with
the weights of NIs fixed). The backpropagated signals at J
are subtracted from the output signals of the NCs to produce
other error signals at L.
ii. These error signals at L are then used to update the weights
of the NCs using the backpropagation algorithm. This
causes the NCs to change its output in a way driving the
error signals at L to zero.
iii. New control signals are calculated ∆ Epd , ∆ Epq for the
shunt branch and ∆Ed, ∆Eq for the series branch using the
updated weights in step ii and are then applied at next time
step (t+1) to the plant at B.
iv. These steps are repeated for the subsequent time periods [4].
D
TDL
Series
Series
Neurocontroller
∆ u (t ) =
∆ Ed ,
SENC
UPFC
Fig.14 SMIB with a single transmission line and a UPFC.
A. Neuroidentification of Plant Dynamics
Identification of the error signals Perr (t), Qerr (t) by SENI
and Verr(t), Vdcerr(t) by SHNI are carried out at different
operating points and their weights are updated continually. The
training signals fed to NIs - ∆Ed_prbs, ∆Eq_prbs and ∆Epd_prbs,
∆Epq_prbs are those shown in Figs. 5 and 7. Figs. 15 and 16
show the outputs of SENI( Pˆ (t ) , Qˆ (t ) ) and Figs. 17 and
Error
err
C
TDL
TDL
Error
∆ uˆ (t )
err
Series
Series
Neuroidentifier
Neuroidentifier
SENI
J
+0.5055
H
+0.5037
+0.5031
SHNC
181.7
Verr (t + 1),
K Vdcerr (t + 1)
∆ u (t ) =
∆ E pd ,
Error
∆ E pq
Vˆerr (t + 1),
Vˆ
(t + 1)
dcerr
L
TDL
TDL
Error
Error
∆ uˆ (t )
Series
Neuroidentifier
J
H
182.1
182.2
+0.1045
+0.1037
+0.1029
+0.1021
Qerr
Q̂err
+0.1013
+0.1005
181.7
181.9
182
Time (sec)
181.8
182.1
182.2
Fig. 16 Actual signal Qerr of the plant and estimated signal Q̂err by the SENI.
Neuroidentifier
+0.075
SHNI
Verr
+0.025
Fig. 13 Post-training of SENC to update the weights of the NI by
backpropagating the error signals at L.
V.
181.9
182
Time (sec)
E
Shunt
Voltage (KV)
C
181.8
Fig. 15 Actual signal Perr of the plant and estimated signal P̂err by the SENI.
Desired
Desired
Response
Response
Predictor
Predictor
TDL
Shunt
Series
Neurocontroller
P̂err
+0.5043
Reactive Power (pu)
A
PLANT
PLANT
B
TDL
Perr
+0.5049
Fig. 12 Post-training of SENC to update the weights of the NI by
backpropagating the error signals at L.
D
err
18 show the outputs of SHNI ( Vˆerr (t ) , Vˆdcerr(t) ).
Pˆ err (t + 1),
(t + 1)
Qˆ
E
Z2
Z1
Infinite
Bus
Desired
Response
Predictor
P err (t + 1),
Q err (t + 1)
K
∆Eq
L
Z1
Real Power (pu)
TDL
A
PLANT
B
2
1
SIMULATION RESULTS
The system model in Fig. 14 comprises of a synchronous
generator (590 MVA, 38 KV L-L) [8] operating at real power,
P=0.5 p.u and reactive power, Q =0.1 p.u, with a transmission
line impedance of Z = (0.02+ j0.4) p.u. The governor and
turbine models are the IEEE standard models of
PSCAD/EMTDC [9]. The parameters of PI controllers are fine
Vˆerr
-0.025
-0.075
-0.125
183.1
183.2
183.3
183.4
183.5
183.5
Fig. 17 Actual signal Vdcerr of the plant and estimated signal Vˆdcerr by the
SHNI.
313
Voltage (KV)
-0.6
load angle responses respectively. For this operating point, it
can be seen that the responses of NCs are found to perform
better than the PI controllers in damping the system
oscillations. The maximum overshoot and the settling time are
less compared to that with the PI controllers. In addition, it can
be observed especially from Figs. 20 and 22 that the PI
controllers’ performances have degraded.
-1
-1.4
-1.8
Vdcerr
-2.2
Vˆdcerr
-2.6 183.1
183.2
183.3
183.4
183.5
183.5
Time (sec)
40
Fig. 18 Actual signal Verr of the plant and estimated signal Vˆerr by the
35
Terminal Voltage (KV)
SHNI.
B. Neurocontrol of the Plant
The NCs and the PI controllers’ performances are evaluated
by applying a 180 ms three phase short circuit fault at bus 2 at
three different operating points given below. The figures below
show the response of the plant with the NCs (SENC and
SHNC) in solid lines and with the PI controllers in dashed
lines.
30
25
20
15
10
NC
5
PI
6
7
8
Time (sec)
9
10
11
Fig. 21 Terminal voltage response of the synchronous generator operating (P =
0.65 pu and Q = 0.12 pu) for a 180 ms three phase short circuit at bus 2.
1) First operating point – P = 0.5 pu and Q = 0.1 pu:
Figs. 19 and 20 show the terminal voltage and speed responses
for the two types of controllers (NCs and PIs) respectively. It
can be observed from these figures that the performance of
both controllers is similar at this operating point. This is
because the PI controllers are fine tuned initially for this
operating point.
1.015
1.01
Speed (Pu)
1.005
1
0.995
40
0.99
NC
0.985
PI
30
5
6
7
25
8
Time (sec)
9
10
11
Fig. 22 Speed response of the synchronous generator operating (P = 0.65 pu
and Q = 0.12 pu) for a 180 ms three phase short circuit at bus 2.
20
15
10
5
NC
0
PI
5.5
6
6.5
7
7.5
8
Time (sec)
8.5
9
120
100
9.5
Load Angle(°)
Terminal Voltage (KV)
35
Fig. 19 Terminal voltage response of the synchronous generator operating (P =
0.5 pu and Q = 0.1 pu) for a 180 ms three phase short circuit at bus 2.
80
60
40
20
1.01
PI
6
1.005
Speed (Pu)
UPFC
NC
0
7
8
Time (sec)
9
10
11
Fig. 23 Load angle response of the synchronous generator operating (P = 0.65
pu and Q = 0.12 p.u) for a 180 ms three phase short circuit at bus 2.
1
0.995
3) Third operating point – P = 0.8 pu and Q = 0.15 pu: Figs.
24, 25 and 26 show the terminal voltage, speed and load angle
responses respectively for this operating point which is much
further away from one at which the PI controllers were fine
tuned. It can be clearly seen from these figures that the plant
with PI controllers have sustained oscillations in the terminal
voltage and the load angle increases drastically after the fault,
loosing stability. The plant with NCs on the other hand damps
out the oscillations and restores the system to stability. The
NC
0.99
PI
6
7
8
9
Time (sec)
10
11
Fig. 20 Speed response of the synchronous generator operating (P = 0.5 pu and
Q = 0.1 pu) for a 180 ms three phase short circuit at bus 2.
2)
Second operating point – P = 0.65 pu and Q = 0.12
pu: Figs. 21, 22 and 23 show the terminal voltage, speed and
314
NCs give performances similar to those at previous operating
points and maintains plant stability. This is because NCs are
trained online and hence they able adapt to changes in
operating conditions with the aid of the neuroidentifiers.
UPFC better than the conventional PI controllers. The superior
performance of the neurocontrollers is expected as a result of
the online training of neuroidentifiers and the neurocontrollers
which never stops. Future work involves extending the control
strategy to a large power system with multi-generators and
studying the interactions between different FACTs devices on
the system.
40
Terminal Voltage (KV)
35
VII. ACKNOWLEDGMENT
30
Financial support from the National Science Foundation
(NSF), USA under the Grant No. ECS-0231632 for this
research is gratefully acknowledged.
25
20
15
10
REFERENCES
NC
5
PI
6
7
8
9
Time (sec)
10
[1]
R.M. Mathur, R. K. Varma, Thyristor-based FACTS controllers for
electrical transmission systems, IEEE Press and John Wiley and Sons,
Inc.
[2] L. Chunlei; S. Hongbo, D.C. Yu, “A novel method of power flow analysis
with unified power flow controller (UPFC)” IEEE Power Engineering
Society Winter Meeting, vol.: 4, pp. 2800 -2805.
[3] N. G. Hingorani, L. Gyugyi, Understanding FACTS concepts and
technology of flexible AC transmission systems, Power Electronics
Sponsored by IEEE Power Engineering Society.
[4] G.K. Venayagamoorthy, R.G. Harley, "Two separate continually onlinetrained neurocontrollers for excitation and turbine control of a turbo
generator", IEEE Transactions on Industry Applications, vol. 38 no. 3, pp.
887 -893, May/June 2002.
[5] G.K. Venayagamoorthy, R.G. Harley, “A continually online trained
neurocontroller for excitation and turbine control of a turbo generator”,
IEEE Trans. on Energy Conversion, vol.16, no.3, pp. 261-269,
September 2001.
[6] P.K. Dash, S. Mishra, G. Panda, “A radial basis function neural network
controller for UPFC”, IEEE Transactions on Power Systems, vol. 15, no.
4, pp. 1293 – 1299, November 2000.
[7] L.Y. Dong, L. Zhang, M.L. Crow, “A new control strategy for the unified
power flow controller”, IEEE 2002 Power Engineering Society Winter
Meeting, vol. 1, pp. 562 -566.
[8] P.M.Anderson, A.A.Fouad, Power System Control and Stability, IEEE
Press, New York, 1994, ISBN 0-7803-1029-2.
[9] Manitoba HVDC Research Centre Inc, PSCAD/EMTDC manual
Version 3.0, 244 Cree Crescent, Winnipeg, Manitoba, Canada R3J 3W1
[10] K. Ogata, Modern control engineering (3rd ed.), Prentice-Hall, Inc.,
Upper Saddle River, NJ, 1996.
11
Fig. 24 Terminal voltage response of the synchronous generator operating (P =
0.8 pu and Q = 0.15 pu) for a 180 ms three phase short circuit at bus 2.
1.1
1.08
Speed (Pu)
1.06
1.04
1.02
1
0.98
NC
0.96
PI
4
5
6
7
8
Time (sec)
9
10
11
Fig. 25 Speed response of the synchronous generator operating (P = 0.8 pu and
Q = 0.15 pu) for a 180 ms three phase short circuit at bus 2.
250
200
Load Angle(°)
150
100
50
0
NC
-50
PI
5
6
7
8
Time (sec)
9
10
11
Fig. 26 Load angle response of the synchronous generator operating (P = 0.8
pu and Q = 0.15 pu) for a 180 ms three phase short circuit at bus 2.
VI.
CONCLUSION
In this paper, the design of two continually online trained
neurocontrollers to provide adaptive nonlinear control for the
series and shunt UPFC branches over a wide range of
operating conditions is proposed. It has been shown here that
it is possible for two neural networks to identify the hybrid
complex dynamics of a unified power flow controller and the
power system and another two neural networks to control the
315